Comfiest

com·fi·est adjective
informal : the superlative form of comfy —
the most comfortable, cozy, or providing the greatest degree of physical ease and relaxation.
comfy noun : ComfyUI

A local-first, open-weights video production suite for ComfyUI — driven by your agent.

Comfie.st · built by byMAR.CO

Website · Start here · Quick start · Workflows · Requirements · Contributing

New in Comfiest

NEW — MiniMax H3 just leveled up. Comfiest ships the latest H3 innovations as runnable graphs and
agent skills, not just links:

• Pika H3 Turbo LoRAs — the official Pika H3 Flex adapters
  from Pika-Labs/Pika-H3-Flex at the 6-step LCM
  contract, 24 fps, with text, first/last-frame, and reference modes.
  Pika T2V
• Upscale refiner LoRA — the Regenerate-2K refiner LoRA
  plus latent 2× upscale for 2560×1440 finals. Upscale workflows
• Looping videos — seamless 154-frame H3 loops with synchronized audio and a ×2 seam check,
  built on the MiniMax H3 base.
  Perfect loop

These are brand-new innovations in the MiniMax H3 ecosystem, and Comfiest supports all three.

Comfiest packages the parts of an AI production pipeline that normally live in someone's head:
model routing, workflow templates, hardware limits, submission loops, output provenance, and the
small traps that waste a render. Point a coding agent at this repository and it can set up the MCP
bridge, pick the right model for a shot, run it, and collect the result.

Built for the Comfy Org Open Source Higgsfield challenge.

No partner nodes. No API keys. No credits. Everything runs on your own ComfyUI with open weights.

Start here: point your agent at the repo

The primary interface is an agent. Open this repository in Claude Code, OpenCode, Codex, Cursor, or
any client with MCP support, and say:

“Set up Comfiest for my ComfyUI server, then run the fox demo and show me the output.”

The agent should:

1. Read README.md, COMFY-SETUP.md, and skills/SKILL.md.
2. If ComfyUI is not installed, follow COMFY-SETUP.md to install ComfyUI, the node packs, and the models.
3. Install/register the comfy-* MCP server for your ComfyUI host using MCP-SETUP.md.
4. Confirm the target with server_info() and /system_stats.
5. Pick a workflow from the routing table and submit it with run_workflow.
6. Collect the output into work/outputs/<project>/<date>/<prompt_id>/ and verify its size.

Add the skills to your agent

If your agent supports skills:

• Copy skills/SKILL.md to <client-skills>/comfiest/SKILL.md for the umbrella rules and routing.
• Copy skills/minimaxh3/, skills/minimaxh3-pika/, skills/minimaxh3-character-replace/,
  skills/minimaxh3-upscale/, skills/minimaxh3-perfect-loop/, skills/ltx2-5/, skills/relayvsr/,
  skills/sam3/, skills/prompt-writing/, skills/workflow-development/, and
  skills/fish-audio-drama3/ as sibling skill directories if your client discovers each skill
  separately. skills/prompt-writing/ contains seven family skills; copy its child directories
  individually if your client requires a flat layout.

The umbrella carries the hard rules and routing; the subskills carry family and recipe detail.

What your agent gets

• Routing that respects hardware. Krea/Flux/MOSS for images, YuE2/Music 3 for songs, Drama 3 for
  speech, MiniMax H3/Pika for video, 2K refine, and seamless loops. Heavy models are called out before
  they blow up a 24 GB card.
• Workflow templates that actually run. UI files for ComfyUI, pre-slimmed api/*.api.json
  mirrors for MCP execution, and conversion notes for the frontend-only nodes that break API runs.
• A development loop for new workflows. Scratch copies, UI-to-API conversion, target preflight,
  submit/poll/collect, run-log entries, and a promotion checklist in
  skills/workflow-development/SKILL.md.
• A native matte primitive. SAM3 image/video segmentation with open-vocabulary prompts, coverage
  checks, and alpha-WebM compositing — no character-replacement graph required.
• A second video family. LTX-2.5 22B distilled generates video and audio in one pass from text,
  an identity reference, or a first frame.
• Prompt-writing skills for every family. Krea, Flux, MOSS, LTX, YuE2, Music 3, and Drama 3
  prompts get the model-specific structure instead of a one-line caption.
• A submission loop that survives real life. wait: false + /history polling, direct output
  URL fallback when local job state lags, one-job-per-target discipline.
• Provenance by default. Every file lands under a prompt_id folder with a run-log entry; vision
  QA is allowed only on your own outputs.
• A writable workspace. work/ is the only place agents create files; the shipped workflows stay
  intact.

Quick start

1. Set up ComfyUI

Install ComfyUI 0.43 or newer with the node packs and models listed in
skills/SKILL.md. A 24 GB-class GPU is the baseline. Drama 3
loads 38.6 GB of experts into system RAM; use a high-RAM host.

For a complete, copy-pasteable install — ComfyUI, all 18 custom-node packs, every model, and the
TensorRT depth engine — follow COMFY-SETUP.md. It includes two scripts:

python scripts/install_nodes.py --comfyui "$COMFYUI"
python scripts/download_models.py --comfyui "$COMFYUI"

2. Connect the MCP bridge

Comfiest talks to ComfyUI through comfy-mcp, a thin stdio
server around comfy-cli.

# macOS / Linux
python -m venv ~/venvs/comfiest
~/venvs/comfiest/bin/pip install comfy-cli comfy-mcp

# Windows PowerShell
python -m venv $env:USERPROFILE\venvs\comfiest
& "$env:USERPROFILE\venvs\comfiest\Scripts\pip" install comfy-cli comfy-mcp

Then register one server per ComfyUI target. Full instructions, environment variables, and client
snippets for OpenCode, Claude Code/Desktop, Codex, Cursor, Hermes, and a generic stdio client are in
MCP-SETUP.md.

3. Run your first workflow

Ask the agent for the fox demo, or submit directly:

const run = await tools["comfy-<host>-<port>"].run_workflow({
  workflow_path: "<repo>/workflows/text2image-krea.json",
  wait: false
});
await tools["comfy-<host>-<port>"].job({ action: "wait", prompt_id: run.prompt_id });
await tools["comfy-<host>-<port>"].fetch_outputs({
  prompt_id: run.prompt_id,
  out_dir: "<repo>/work/outputs/fox-demo/<date>/" + run.prompt_id
});

Replace <repo> with the absolute path to this repository on the MCP host, and <host>/<port> with
one of your registered servers.

Demo inputs

The demo prompts are self-contained, but the reference-driven workflows expect files on the ComfyUI
host:

Voice-clone inputs: The transcripts in the TTS templates are placeholders (“The quick brown
fox…”). A clone only works when the transcript matches the clip word for word, so replace both.

Workflow catalog

MiniMax H3 is the video workhorse: text-to-video, 1–2 picture references, video/audio references,
forced-audio lip-sync, and character replacement. See
skills/minimaxh3/SKILL.md for the variant matrix and
minimaxh3-character-replace for the swap pipeline,
minimaxh3-upscale for 2K t2v/v2v refine, and
minimaxh3-perfect-loop for seamless 154-frame loops.

Before filling any prompt field, use the matching prompt writer from
skills/prompt-writing/. They cover Krea, Flux, MOSS, LTX, YuE2,
Music 3, and Drama 3; the index maps each to its workflow.

Repository layout

.
├── README.md                # this file
├── LICENSE                  # MIT (repository); model weights keep their upstream licenses
├── COMFY-SETUP.md           # install ComfyUI, nodes, models, TensorRT engine
├── MCP-SETUP.md             # install, register, verify comfy-mcp
├── assets/
│   ├── icon.png             # site icon (cat mark, used in the header link)
│   └── logo.png             # project logo
├── skills/
│   ├── SKILL.md             # umbrella: rules, routing, operating loop
│   ├── minimaxh3/SKILL.md   # fastvideo variant matrix and execution notes
│   ├── minimaxh3-pika/SKILL.md  # Pika H3 Flex 6-step default
│   ├── minimaxh3-character-replace/SKILL.md
│   ├── minimaxh3-upscale/SKILL.md
│   ├── minimaxh3-perfect-loop/SKILL.md
│   ├── ltx2-5/SKILL.md                # LTX-2.5 video + joint audio
│   ├── relayvsr/SKILL.md              # RelayVSR X2 video upscale
│   ├── sam3/SKILL.md                  # native masking family
│   ├── prompt-writing/                # seven family prompt writers + README index
│   ├── workflow-development/SKILL.md   # edit -> validate -> run -> promote
│   └── fish-audio-drama3/SKILL.md
├── scripts/
│   ├── install_nodes.py     # clone/update custom nodes, pip requirements
│   ├── download_models.py   # Hugging Face downloads into ComfyUI folders
│   ├── validate_workflow.py # check an API graph against a live target
│   ├── wait_job.py          # poll /history when the MCP wait fails
│   └── fetch_output.py      # download outputs when fetch_outputs fails
├── workflows/               # 62 workflow templates
│   ├── *.json
│   ├── fishdrama3/          # TTS base + variants
│   ├── ltx/                 # LTX-2.5 t2v/ref2v/i2v graphs
│   ├── relayvsr/            # RelayVSR X2 video upscale
│   ├── sam3/                # native image/video matte graphs
│   └── minimaxh3/           # H3 base, UI variants, api/ mirrors
└── work/                    # writable agent workspace (not required to run)
    ├── inbox/  scratch/  uploads/  outputs/
    ├── analysis/  post/  exports/  logs/
    └── README.md

Design principles

1. Open weights only. Every workflow is a local ComfyUI graph. No partner/cloud nodes, no API
   keys, no credits.
2. Agent-native. The skill is written for a machine reader: explicit rules, exact filenames, node
   IDs, known failure modes, and a provenance contract.
3. Templates, not lock-in. Copy a workflow to work/scratch/, change it, run it. The shipped
   graphs are a starting point.
4. One job per target. Parallel work only across different ComfyUI hosts.
5. Provenance over vibes. A prompt_id, a run-log row, and a non-zero file size prove a run
   happened. Vision is QA, not verification.
6. Stable save prefixes. Never rewrite a workflow's save prefix; identify and collect runs by
   prompt_id. The shared ComfyUI output/ tree is not a place to invent folder names.

Public-release sanitization

This repository is the sanitized public edition of a working production suite:

• Non-functional LoRAs (character, style, anatomy, adult-oriented) were removed. The surviving LoRAs
  are functional: MiniMax H3 turbo adapters (v1.2 and 4–8 step), Pika H3 Flex fl2va/ref2va, the 2K
  refiner and LMS sharpener, the LTX Best-FaceID identity LoRA, and the MOSS-Image DMD sampler.
• Every baked-in prompt was replaced with a neutral demo prompt (fox-themed for t2v). Reference
  image, video, and audio names were genericized to example.*.
• UI workflows were conversion-validated against a live ComfyUI target; the sanitized graphs have not
  been executed end-to-end. Run one job before relying on a template in production.

Add your own LoRAs and prompts in work/scratch/. If a look needs a custom LoRA, document it in your
fork rather than editing the shipped template.

Requirements

• ComfyUI 0.43+ and frontend 1.57.1+.
• comfy-cli + comfy-mcp (Python 3.12 tested).
• Node packs: the full set is installed by COMFY-SETUP.md — rgthree, RES4LYF,
  WAS, Easy-Use, various, KJNodes, VideoHelperSuite, Drama3, krea2edit, Conditioning Rebalance,
  PlagueKind, Context-Loop, BFSNodes, RelayVSR, Fill-Nodes, Spectrum, Impact Pack, and
  Depth-Anything-TensorRT.
  Optional: Frame-Interpolation (the H3 graphs use Fill-Nodes FL_RIFE), the latent upscaler, and
  ComfyUI-GGUF (legacy GGUF graphs only).
• Models: Krea 2 Turbo/Edit, Flux.2 Klein 9B + Qwen3-8B fp8mixed, MOSS-Image, YuE2 3B, MiniMax Music 3,
  Fish Audio Drama 3, MiniMax H3 fastvideo stack + Pika H3 Flex adapters + 2K refiner/upscaler and the
  loop Turbo adapter, LTX-2.5 22B distilled + Gemma 4 12B + Best-FaceID identity LoRA, RelayVSR X2,
  SAM3 segmentation. Exact filenames live in the workflow sections of
  skills/SKILL.md; verify each host with /object_info/<LoaderClass> option
  lists. /models/<folder> can return an empty body even when the model is installed.
  The full download list, Hugging Face locations, and folder layout are in
  COMFY-SETUP.md.
• GPU: 24 GB-class for video and Drama 3; 128 GB+ system RAM recommended for Drama 3.

Model weights are not distributed here. Download them from their upstream publishers and follow their
licenses.

Contributing

Issues and PRs are welcome, especially new workflows, node-pack compatibility fixes, and agent
ergonomics.

• Keep workflows runnable: include both the UI file and, where useful, an API mirror under
  workflows/<family>/api/.
• Update skills/SKILL.md and the family README with inputs, settings, save prefix, and known traps.
• Run comfy run --workflow <file> --print-prompt --host <host> --port <port> before submitting.
• Never change a workflow's save prefix (filename_prefix) — retrieval is by prompt_id.
• Never commit work/ contents, model weights, generated outputs, or credentials.

License

Repository code and documentation: MIT.

Workflow JSON is provided under the same MIT terms. Model weights, LoRAs, and node packs remain under
their own upstream licenses — check each model card before commercial use. “Comfiest” is not
affiliated with Comfy Org, Krea, Black Forest Labs, MiniMax, Fish Audio, or YuE2.

Acknowledgments

ComfyUI and comfy-mcp by Comfy Org; Krea 2, Flux.2 Klein, YuE2, MiniMax H3/Music 3, and Fish Audio
Drama 3 by their respective teams; rgthree, RES4LYF, KJNodes, VideoHelperSuite, Frame Interpolation,
and the other node-pack authors who make these graphs possible.